Prabhjotschugh/firstpass-peer-review

FirstPass

FirstPass is a multi-domain, multi-round scientific peer-review dataset built from Nature Communications transparent peer-review files. It contains 3,668 complete peer-review dialogues across five scientific domains, with editorial outcome labels derived from real revision-cycle outcomes.

The dataset was introduced alongside a fine-tuned model (Qwen2.5-7B-Instruct + LoRA) that achieves 80.5% accuracy and F1-macro 78.2% on revision-cycle prediction (STANDARD vs. EXTENDED), and was accepted to the AI Scientists: Tools, Co-authors, or Founders? workshop at ICML 2026.

Note: The accompanying paper has been accepted to the AI4Science Workshop at ICML 2026. A separate dataset-track submission is under review. Full citation details and model weights will be added after final results.

---

Dataset Summary

Property	Value
Total records	3,668
Domains	Biology, Chemistry, Neuroscience, Physics, Earth Science
Source journal	Nature Communications (ISSN 2041-1723)
Date range	January 2023 – December 2025
License	CC BY 4.0 (inherited from source)
Content integrity	100% verified (zero hollow files)
Avg. review length	2,155 words
Label balance	65.4% STANDARD / 34.6% EXTENDED

Domain breakdown

Domain	Records	STANDARD (%)	EXTENDED (%)	Avg Review (words)
Biology	741	63.2	36.8	2,201
Chemistry	744	62.8	37.2	2,183
Neuroscience	739	64.6	35.4	2,144
Physics	727	66.7	33.3	2,098
Earth Science	717	71.2	28.8	2,149

---

Configs / Subsets

The dataset has three configs corresponding to different stages of the pipeline:

cls — Classification splits (outcome prediction)

Used to train and evaluate the revision-cycle prediction task.

from datasets import load_dataset
ds = load_dataset("Prabhjotschugh/firstpass-peer-review", "cls")
# ds["train"], ds["validation"], ds["test"]

Each example:

{
  "task": "outcome_prediction",
  "input": "<paper text + review dialogue>",
  "target": "STANDARD",
  "article_id": "s41467-024-...",
  "domain": "biology",
  "n_rounds": 2,
  "label_src": "heuristic"
}

Labels: STANDARD (2-round revision cycle) · EXTENDED (3+ round revision cycle)

Split	Examples
train	~2,338
validation	~294
test	318

sft — Generation splits (review generation + reviewer updating)

Used for supervised fine-tuning of the generative tasks.

ds = load_dataset("Prabhjotschugh/firstpass-peer-review", "sft")

Each example:

{
  "task": "review_generation",   // or "reviewer_updating"
  "input": "<paper text>",
  "target": "<full verbatim reviewer comments>",
  "article_id": "s41467-024-...",
  "domain": "chemistry",
  "n_rounds": 2,
  "label_src": "heuristic"
}

Tasks in SFT:

• review_generation — paper → Round 1 reviewer comments
• reviewer_updating — paper + Round 1 reviews + author response → Round 2 reviewer comments

Split	Examples (approx.)
train	~5,500
validation	~690
test	~740

records — Full structured records (one per paper)

The complete parsed records with all peer-review rounds, paper content sections, training examples, and metadata. Useful for custom preprocessing or analysis.

ds = load_dataset("Prabhjotschugh/firstpass-peer-review", "records", split="biology")

Each record contains:

• doi, article_id, title, pub_date, domain
• paper_content: {abstract, introduction, results, methods}
• peer_review_rounds: list of {round, reviewer_comments, author_response}
• decision_letter, n_rounds, total_reviewer_words
• outcome_label: major_revision / minor_revision / accept / unknown
• training_examples: pre-built input/target pairs for all three tasks
• _stats: per-record diagnostics (chars, sections found, rounds summary, etc.)

---

Data Collection

1. Discovery: Queried the Springer Nature OpenAccess API (ISSN 2041-1723) for Nature Communications articles published 2023–2025 across five domain-specific subject filters.

2. Download: Scraped article landing pages on nature.com to retrieve both the paper PDF and the transparent peer-review PDF (mandatory at Nature Communications since November 2022).

3. Parsing: Both PDFs are processed through Gemini-3.1-flash-lite-preview using strongly-engineered extraction prompts with a six-layer JSON recovery pipeline. The paper parser extracts four sections (abstract, introduction, results, methods). The review parser segments reviewer comments, author responses, and decision letters by revision round.

4. Quality filtering: A record is retained only if:

   • All four paper sections exceed 20 words each
   • At least two complete review rounds are present
   • Total reviewer word count ≥ 150

5. Integrity audit: Automated audit confirmed 0 hollow files across all 3,668 records (100% content integrity).

6. Splitting: 80/10/10 paper-level split, stratified by (domain, label) to ensure balanced class and domain distributions in every split. Random seed: 42.

---

Tasks and Use Cases

Task 1 — Review Generation

Input: Paper (abstract + introduction + methods + results)
Target: Round 1 reviewer comments (verbatim, all reviewers concatenated)
Use case: Train a model to generate expert-level peer reviews from a manuscript.

Task 2 — Reviewer Updating

Input: Paper + Round 1 reviews + author point-by-point response
Target: Round 2 reviewer comments
Use case: Train a model to update a review after reading the author rebuttal, identifying which concerns were resolved and which remain outstanding.

Task 3 — Outcome Prediction (primary evaluation task)

Input: Full dialogue (paper + all rounds of reviews + author responses)
Target: STANDARD or EXTENDED
Use case: Predict whether a manuscript will require a standard (2-round) or extended (3+ round) revision cycle, enabling pre-submission anticipatory critique.

---

Key Findings (from the companion paper)

• Response-only loss masking is a prerequisite, not an optimization. Without it, fine-tuned accuracy collapses to 62.0% (below the 65.4% majority baseline). With it, the model achieves 80.5% accuracy and F1-macro 78.2%.
• Cross-domain generalization: Per-domain accuracy ranges from 76.9% (Earth Science) to 83.8% (Biology), a 6.9 pp spread, confirming the model captures domain-general signals of unresolved reviewer concern.
• Review generation: The fine-tuned model produces reviews averaging 1,187 words, significantly closer to human references (2,155 words) than any zero-shot baseline, achieving ROUGE-L 0.154.

---

Fine-tuning Recipe (recommended)

# Classification (CLS adapter)
# Base: Qwen2.5-7B-Instruct
# LoRA: r=32, alpha=64, target all 7 projection matrices
# CRITICAL: response-only loss masking (train_on_responses_only)
# Epochs: 3, max_seq_len: 12288, lr: 5e-5, batch: 16 (effective)
# Optimizer: paged_adamw_8bit, scheduler: cosine, warmup: 30 steps

# Generation (SFT adapter)
# Same LoRA config, 1 epoch, max_seq_len: 16384
# Repetition penalty 1.1 at inference

See the companion paper for full hyperparameters and ablation results.

---

Citation

Note: Paper citation details will be updated after final workshop proceedings are published. If you use this dataset before then, please cite it as:

@misc{prabhjot_singh_2026,
    author       = { Prabhjot Singh and Somnath Luitel and Manmeet Singh and Josh Durkee },
    title        = { firstpass-peer-review (Revision 1aac536) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/Prabhjotschugh/firstpass-peer-review },
    doi          = { 10.57967/hf/8982 },
    publisher    = { Hugging Face }
}

---

License

The dataset is released under CC BY 4.0, consistent with the Creative Commons license of the source Nature Communications peer-review files.

---

Contact

For questions about the dataset, please open a Discussion on this repository.